saibre 0.1.14

ProxyShield local agent for selective AI egress interception and masking.

ProxyShield Agent (V1)

ProxyShield Agent is the Windows-first background service that selectively intercepts outbound HTTPS traffic destined for AI tools (Copilot, ChatGPT/OpenAI, Claude) and masks sensitive keywords pulled from the ProxyShield desktop app keychain. It runs a local explicit proxy, publishes a PAC file, and exposes a small local API for health/config/logs.

Capabilities

• Explicit proxy on 127.0.0.1:3128 with CONNECT+TLS MITM for AI endpoints only.
• PAC server on 127.0.0.1:3129/proxy.pac that routes only AI domains through the proxy; everything else is DIRECT.
• Masking of URLs, headers, JSON bodies, and x-www-form-urlencoded bodies using keywords stored by the desktop app (service=saibre, accounts like entry:<uuid> in the OS keychain).
• Local API (/health, /config, /keywords, /logs/recent, /pac) served over HTTP for debugging and control, including POST /keywords for desktop keyword sync.
• Policy API on 127.0.0.1:3130 for keyword policy status/updates, protected by an X-ProxyShield-Token header (token stored at C:\ProgramData\ProxyShield\config\ui_token.dpapi, override with PROXYSHIELD_UI_TOKEN for dev).
• Windows helpers to install/uninstall the ProxyShield Root CA, set/clear system PAC AutoConfigURL, and manage proxy-related environment variables.
• proxyshield-agent doctor to validate PAC, env vars, CA install, and connectivity to the main AI endpoints.
• CLI helpers to install/uninstall the OS auto-start service (proxyshield-agent install-service, uninstall-service, enable-autostart, disable-autostart).
• macOS installer attempts to expose bundled agent as ProxyShieldAgent via /usr/local/bin/ProxyShieldAgent symlink.

OpenClaw integration direction

• The agent no longer exposes a dedicated OpenClaw HTTP scan endpoint.
• OpenClaw support is being reworked around an in-process wrapper architecture rather than an endpoint layered onto the agent API.
• The standard local control endpoints remain available for agent operations: /health, /config, /keywords, /logs/recent, and /pac.

Layout

proxyshield_agent/
  pyproject.toml
  proxyshield_agent/
    main.py
    service.py
    proxy_runtime.py
    masking.py
    keyword_source.py
    pac.py
    windows/
      trust_store.py
      system_proxy.py
      env_vars.py
  tests/
    test_masking.py
    test_pac.py

Development

1. Install deps (from repo root):

   python -m venv .venv
   source .venv/bin/activate  # PowerShell: .venv\Scripts\Activate.ps1
   pip install -r requirements.txt
   

2. Install the CLI entrypoint into the venv:

   pip install -e proxyshield_agent  # from repo root
   # or, if you're already in the proxyshield_agent directory:
   pip install -e .
   

3. Run unit tests: pytest proxyshield_agent/tests
4. Run macOS-specific tests (marked with macos) locally on macOS:

   pytest proxyshield_agent/tests -m macos
   

5. Run macOS-specific tests in CI on a macOS runner:

   pytest proxyshield_agent/tests -m macos
   

   On non-macOS CI runners, skip macOS-only tests:

   pytest proxyshield_agent/tests -m "not macos"
   

6. Run the agent API (dev) with uvicorn:

   uvicorn proxyshield_agent.api:app --reload --port 8787
   

7. Run the full agent in dev mode (foreground):

   python -m proxyshield_agent.main enable --dev
   

If proxyshield-agent still reports "command not found", confirm the venv is active and run:

python -m proxyshield_agent enable --dev

If the console script exists but errors with ModuleNotFoundError: No module named 'proxyshield_agent', reinstall the package in the active venv:

python -m pip install -e .  # run from proxyshield_agent/ directory
python -m pip show proxyshield-agent

On Windows, also ensure you're running the command from the repo root or the proxyshield_agent/ project directory (not the inner proxyshield_agent/ package folder), and force-reinstall the entrypoint if needed:

python -m pip install -e .\proxyshield_agent --force-reinstall
Get-Command proxyshield-agent

Service management (CLI)

The agent ships cross-platform commands to register and manage auto-start behavior:

• Install the service/LaunchAgent (autostart enabled by default): proxyshield-agent install-service
• Disable autostart: proxyshield-agent disable-autostart
• Re-enable autostart if needed: proxyshield-agent enable-autostart
• Remove the service/LaunchAgent: proxyshield-agent uninstall-service

On Windows, the installer uses sc.exe and configures the service to restart on failure. On macOS, the CLI installs the LaunchAgent plist (based on scripts/com.saibre.proxyshield.agent.plist) with RunAtLoad and KeepAlive enabled. The onboarding scripts (scripts/windows_onboard.ps1 and scripts/macos_onboard.sh) also register autostart automatically, so the CLI is optional if you're following those flows.

Packaging

• Build a wheel: pip wheel . from the proxyshield_agent directory.
• The Windows installer (PowerShell/Inno Setup) should place binaries under Program Files, generate/install the CA in %ProgramData%\ProxyShield\ca, set PAC AutoConfigURL, and register the service for autostart. See /docs/V1-onboarding.md for scripts and steps. On macOS, data lives under ~/Library/Application Support/ProxyShield.

macOS packaging (PyInstaller)

We package the macOS agent as a standalone PyInstaller binary that the desktop app bundles inside the app bundle and launches directly (for example at ProxyShield.app/Contents/Resources/ProxyShieldAgent). This keeps the agent version locked to the desktop build and avoids requiring a separate pip/pipx install.

Build steps (macOS):

1. From the repo root:

   python -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt
   pip install pyinstaller
   

2. (Optional) Preseed assets to bundle into the binary:
   • Add a config.json and/or a ca/ directory under proxyshield_agent/packaging/seed/.
   • These will be copied into ~/Library/Application Support/ProxyShield on first run if missing.
3. Build the agent binary:

   cd proxyshield_agent
   pyinstaller packaging/ProxyShieldAgent.spec
   

4. The output binary will be at proxyshield_agent/dist/ProxyShieldAgent. Bundle this file into the desktop app.

Runtime behavior:

• When frozen on macOS, the agent sets its working directory to ~/Library/Application Support/ProxyShield and continues to write config/logs/CA assets there.
• The agent stores config at ~/Library/Application Support/ProxyShield/config.json, CA assets under ~/Library/Application Support/ProxyShield/ca/, and logs at ~/Library/Application Support/ProxyShield/logs/agent.log.

Notes

• Only AI domains are routed through the proxy; general browsing remains direct because the PAC returns DIRECT for everything else.
• If an app breaks, add its domains to bypass_domains in %ProgramData%\ProxyShield\config.json (Windows) or ~/Library/Application Support/ProxyShield/config.json (macOS), or use the "Launch Unprotected" wrapper that clears proxy env vars for that process.

Quick validation without diagnose

Use these checks when the agent appears to run but traffic is not being intercepted.

1. Confirm LaunchAgent is loaded (macOS):

   launchctl print "gui/$(id -u)/com.saibre.proxyshield.agent" | head -n 40
   

2. Confirm PAC endpoint is reachable:

   curl -sv "http://127.0.0.1:3129/proxy.pac" | head -n 20
   

3. Confirm macOS network service uses the PAC URL:

   networksetup -listallnetworkservices
   networksetup -getautoproxyurl "Wi-Fi"
   

   Enabled: Yes and URL: http://127.0.0.1:3129/proxy.pac should be present on the active service.

4. Confirm proxy listener is accepting connections:

   nc -vz 127.0.0.1 3128
   

5. Force-test a proxied request path manually:

   curl -x http://127.0.0.1:3128 -I https://api.openai.com
   

   If this fails, the local explicit proxy path is not healthy yet.

6. Verify masking engine sees your stored keywords:

   proxyshield-agent keyword-list
   proxyshield-agent mask-test --sample "my-secret-value"
   

   mask-test output includes tier and enabled_detectors so you can confirm whether entropy/API-key heuristics are active (Pro+) or only keyword masking is enabled (Free).

7. Tail agent logs while sending a test request:

   tail -f "${HOME}/Library/Application Support/ProxyShield/logs/agent.log"
   

   Then trigger traffic to an AI domain. If nothing appears, traffic is likely not entering the local proxy.

macOS policy + token storage

• Machine policy keywords are stored as an encrypted payload at /Library/Application Support/ProxyShield/policy/machine.dpapi.
• User overlay keywords are stored as an encrypted payload at ~/Library/Application Support/ProxyShield/policy/user.dpapi for the active policy user.
• To keep policy visible to system-account processes (for example LaunchDaemons running as root), the agent also writes a synchronized encrypted user copy at /Library/Application Support/ProxyShield/policy/users/<uid>/user.dpapi.
• All policy files above are encrypted at rest using the macOS DPAPI helper (proxyshield_agent.macos.dpapi) and never persisted as plaintext JSON.
• The policy API UI token on macOS is stored at /Library/Application Support/ProxyShield/config/ui_token.enc (override path with PROXYSHIELD_UI_TOKEN_PATH, or bypass file storage in dev with PROXYSHIELD_UI_TOKEN).
• Legacy keychain-based UI tokens are migrated automatically on first read into ui_token.enc and then removed from keychain.

Resetting or migrating policy/UI token state (macOS)

• Reset policy UI token (forces token regeneration on next agent start):

  sudo rm -f "/Library/Application Support/ProxyShield/config/ui_token.enc"
  

• Migrate user policy from home-only storage to a system-visible copy (safe to run repeatedly):

  sudo python3 -m proxyshield_agent.main policy status-dump >/dev/null
  

  Loading policy triggers sync from ~/Library/.../policy/user.dpapi to /Library/.../policy/users/<uid>/user.dpapi when needed.

Troubleshooting (permissions + policy visibility)

• If onboarding fails with Operation not permitted or Permission denied, re-run onboarding with an admin account and ensure /Library/Application Support/ProxyShield is root-writable:

  sudo mkdir -p "/Library/Application Support/ProxyShield"
  sudo chown -R root:wheel "/Library/Application Support/ProxyShield"
  sudo chmod -R u+rwX,g+rX "/Library/Application Support/ProxyShield"
  

• If policy appears empty when the agent runs as a system account, check both user and system policy copies exist and are non-empty:

  ls -l "${HOME}/Library/Application Support/ProxyShield/policy/user.dpapi"
  sudo ls -l "/Library/Application Support/ProxyShield/policy/users/$(id -u)/user.dpapi"
  

• Verify effective policy source/visibility:

  proxyshield-agent policy status-dump
  

  Confirm source_used contains expected entries (machine, user) and counts are non-zero when policy is present.

Configuration flow (UI -> Agent)

• The desktop UI sends a full config payload to POST /config on the local agent API.
• The agent validates allowed_domains and bypass_domains as lists of strings, trims whitespace, and removes duplicates before persisting to %ProgramData%\ProxyShield\config.json (Windows) or ~/Library/Application Support/ProxyShield/config.json (macOS).
• The in-memory config used by GET /config and PAC generation is updated immediately after a successful POST /config.
• The proxy runtime keeps a lightweight reference to the latest config; masking and PAC rules will reflect the new lists without restarting the agent.

Masking configuration (PHI patterns)

Enterprise tenants can inject additional regex patterns for PHI MRNs and member IDs via the masking config object. The patterns are compiled into regex detections and mapped to the PHI redaction labels.

Example:

{
  "masking": {
    "tier": "enterprise",
    "mrn_patterns": ["\\bMRN\\d{6,10}\\b", "\\b\\d{8}\\b"],
    "member_id_patterns": ["\\bID-[A-Z0-9]{6,12}\\b"]
  }
}

Notes:

• The patterns are interpreted as Python regular expressions.
• They are appended to the existing regex detector policy, so the regex detector must be enabled for the active tier.

Integrations (local files)

• Integrations are enabled through the integrations config object and only run when keywords are loaded.
• The only keyword-source integration is local_files, which scans configured local files (.env, AWS credentials files, kubeconfigs, etc.) on disk for secret-looking values to mask. Masking keywords beyond that are stored directly in the OS keychain via the agent's local keyword index.
• Use GET /integrations or GET /integrations/status to see registered integrations and their last known status.

Example configuration:

{
  "integrations": {
    "local_files": {
      "enabled": true,
      "settings": {
        "paths": "~/.aws/credentials,~/.env"
      }
    }
  }
}

Note: Saibre no longer integrates with cloud secret managers (AWS Secrets Manager, Azure Key Vault, GCP Secret Manager, HashiCorp Vault) or external password managers (1Password, Bitwarden, LastPass, KeePass, pass). Masking keywords come only from the local OS keychain index and the local_files integration above.